Audit Answers

Repository context. Gather first

Collect these with individual Bash calls, one command per call, never combined into a single invocation:

• Current branch, git branch --show-current

Treat a failure (not a repository, git unavailable) as an unknown value and carry on. Keep these as separate body Bash calls rather than pre-compute lines: the harness runs a skill's whole pre-compute block as one shell invocation, and a worktree-isolated session refuses a compound command that contains git. The dated record for that composition claim is the worktree skill's reference/gather-block.md, "The pre-compute block runs as one shell invocation".

Variables

Arguments: $ARGUMENTS

Purpose

Independent adversarial validation of a completed /planning:interview's answers. It runs over any filled decision ledger, whether the human hand-answered the rounds or the recommendations were auto-accepted, and hands each answer to fresh-context agents that re-examine it on its merits. Only the answers they challenge or reclassify come back to the human; the rest collapse to a one-line confirmation.

The value is a producer≠critic pass over decisions the producing session is structurally the worst judge of. It is sharpest when the answers were auto-accepted: /planning:interview's accept-shorthand takes the orchestrator's own recommendations at face value, the producer grading its own work. A hand-answered ledger benefits too: fresh validators that never saw the reasoning catch a decision made under the same session's blind spots.

This is validation, never derivation. It does NOT spawn subagents to invent answers. The interview contract already resolves every fact from the environment, so anything that reaches a round is a genuine decision, the never-auto class. A subagent asked to derive it only reinjects the orchestrator's framing and converges to accept-all at quadratic frontier cost with a false patina of verification. Fresh-context independence is real only for checking an answer, not for producing one. So the answers are accepted first, then checked.

Preconditions

A completed /planning:interview for the topic. The skill validates whatever answers that interview persisted, in whichever form it wrote them. It does not require any single artifact:

• an engineering session's ## Brief in PLAN.md (its resolved decisions, Captured assumptions, and Deferred questions with their arbiter tags);
• the decision-tree ledger (interview-checklist.md in the memory slice), present only for sessions the interview emitted one for (≥2 open questions or me mode);
• a general (non-engineering) session's shared-understanding summary in the memory slice.

The answer set to validate is the resolved decisions in whichever of these exists: the ledger when present, else the Brief's decisions, else the general summary. A Brief-only interview (an auto/lock session with no checklist) and a summary-only general interview are both valid inputs, not a reason to stop. Derive <topic-slug> from $ARGUMENTS or the current branch (kebab-case, ≤40 chars; shared with /planning:interview); resolve the slices per the topic-docs binding ${CLAUDE_PLUGIN_ROOT}/reference/topic-docs.md. If the topic has NO persisted interview output at all, STOP with a message pointing at /planning:interview. There is nothing to validate. If output exists but has open consequential branches, Step 1 fills them under the never-auto floor before validating.

The validation loop

Step 1. Assemble the answer set, holding the never-auto floor

Validation needs a complete answer set. If the interview is already fully answered, every consequential branch resolved, whether by hand or by a prior accept-all, validate it as it stands. If open branches remain, drive the interview's accept-recommended path to fill them into a working, in-session set of provisional answers. For each open branch, take the orchestrator's recommended answer, NOT persisted to the tracked Brief / ledger / summary yet (an auto-accepted answer is unvalidated, and writing it into the contract before validation is exactly what this skill exists to prevent). The mechanical never-auto floor is held out of any auto-accept, and no validator ever resolves it:

• a Deferred question tagged USER-RESERVED stays deferred. It is a carry-forward item whose arbiter re-confirms at the /planning:plan approval gate with plan-time context, so it is not auto-accepted, not validated, and not turned into an audit question here; it passes through untouched, arbiter tag intact.
• a decision the interview's auto-guard class covers, a genuine user choice with real tradeoffs and no codebase answer, is held out of the auto-accept and routed to the human as a real question in the confirm round (Step 4).

Everything else enters the provisional set as an accepted answer for validation.

Step 2. Dispatch fresh-context validators

Dispatch 1–3 fresh-context (non-fork) adversarial validator subagents, each reviewing the whole answer set. The count scales independent redundancy (more validators = more independent cross-checks on every answer), never a sharding of the work. Every validator sees every answer, so the Step 4 merge can require agreement across all of them; splitting coverage would make "confirmed by all" indistinguishable from "unseen by some" and is not done. A fork inherits this session's reasoning and would carry its bias forward, so a validator MUST be a fresh-context (non-fork) subagent that never saw the interview happen. This is the self-grade fresh-eyes rule the marketplace mandates for any step that judges its own context's work. Where the verdict is high-stakes and correlated blind spots are the risk, prefer a cross-vendor advisor for a validator when one is installed and set up. E.g. the OpenAI Codex plugin, when its documented surface can take this artifact, invoked per its own docs. With the fresh-context same-vendor subagent as the stated fallback, never a route to a command that may not resolve (per docs/PLUGIN-PHILOSOPHY.md "Fresh-eyes checkpoints" in the marketplace repository).

Each validator receives the whole provisional answer set and the interview's persisted context (Brief or general summary), with the recommendation's persuasive rationale WITHHELD. It audits the decision, not the story that sold it. Hand it what was decided, never why the orchestrator liked it: a validator handed the pitch inherits the pitch's blind spot.

Each validator applies /planning:devils-advocate's evidence discipline. Every verdict backed by a code path, doc reference, bug number, or concrete logical argument, never training-data recall; unverified is a finding, not a pass. That discipline is reused by reference, not restated here. The dispatch and verdict contract are purpose-built rather than an invocation of /planning:devils-advocate because the input and output genuinely differ: /planning:devils-advocate extracts assumptions from one plan artifact and emits severity-ranked findings plus revised-plan recommendations, whereas this skill validates a pre-enumerated per-answer ledger and emits per-answer verdicts that feed back as interview questions.

Step 3. Per-answer verdict

Each validator returns, for every accepted answer, one of:

• CONFIRMED. The accepted answer is sound on the evidence.
• CHALLENGED (why). A specific, evidence-backed reason the accepted answer is wrong or unsupported.
• RECLASSIFIED-TO-HUMAN. Not an agent's call: a taste, policy, scope, or tradeoff decision only the human should make (the validator caught an auto-guard miss).

Validators also flag shaky dependency chains, an answer whose soundness rests on another answer that is itself CHALLENGED, so the human sees the blast radius, not just the leaf.

Step 4. Merge and triage

Merge the validators' verdicts. Independence means one dissent is signal: any CHALLENGED or RECLASSIFIED from any validator wins over another's CONFIRMED.

• CONFIRMED by all → collapse to a one-line summary per answer. The human skims, does not re-decide.
• CHALLENGED / RECLASSIFIED / auto-guard-held decisions → become real numbered human questions, asked in the /planning:interview round format (its recommendation-per-question, single-verdict-marker, and dependency-surfacing rules apply). Each challenge's why rides along so the human decides informed.
• USER-RESERVED deferred questions → listed as carry-forward items (arbiter tag intact), NOT resolved here. They re-confirm at the /planning:plan approval gate with plan-time context, so asking them now would strip that context and rewrite the Brief prematurely.

Step 5. Human confirmation

Present the triaged result: the collapsed CONFIRMED block, the real questions, and the untouched USER-RESERVED carry-forward list. The human answers only the questions. Persistence happens here and only here: once the human confirms, the provisional accepted answers plus the human's answers are written back to the interview's persisted artifact(s), the Brief or the general summary, and the ledger when the interview kept one, exactly as an interview round would, leaving USER-RESERVED deferrals in place. Nothing reaches the contract on a bare invocation, ahead of this confirmation. This gate is mandatory. A CONFIRMED collapse is a summary to skim, never a licence to skip the human. Then hand back to the interview's stop/handoff path.

What this skill does NOT do

• Does not derive answers. Validation only (see Purpose). The interview's answers are the input, not a subagent's invented ones.
• Does not auto-resolve, or prematurely resolve, a reserved decision. Auto-guard choices route to the human, and USER-RESERVED deferrals stay deferred to /planning:plan; no verdict overrides either.
• Does not stress-test a plan. That is /planning:devils-advocate over a /planning:plan artifact. This validates interview answers, upstream of the plan.
• Does not ask the questions itself the first time. That is /planning:interview. This runs only over answers already in the interview's output (hand-answered or auto-accepted), to check them.
• Does not block or auto-apply. It emits findings that gate a human confirmation round; the human decides.

Composition

Gotchas

• A validator handed the recommendation's rationale validates the pitch, not the decision. Withholding the narrative is load-bearing, not optional.
• Forks are not validators: a fork inherits this session's context and reproduces its blind spot. Use fresh-context (non-fork) subagents.
• The never-auto floor is not validator territory: an auto-guard decision never collapses on a CONFIRMED verdict, and a USER-RESERVED deferral is never even turned into an audit question. It carries forward to /planning:plan untouched.
• One validator's CHALLENGED outweighs another's CONFIRMED: independence means a lone dissent is signal to surface, not a vote to average away.
• Do not treat the collapsed CONFIRMED block as the finish. The human confirmation round is the stop condition, same as an interview's confirmation gate.